Human pose tracking in low-dimensional subspace using manifold learning by charting

Tracking full articulated human body motion is a very challenging task due to the high dimensionality of human skeleton model, self-occlusion and large variety of body poses. In this work, we explore a novel Low-dimensional Manifold Learning (LDML) approach to overcome high dimensional search space of human model. Low-dimensional demonstration not only delivers a compact tractable search space, but it is efficient to capture general human pose variations. The key contribution of this work is an algorithm of Quantum-behaved Particle Swarm Optimization (QPSO) for pose optimization in latent space of human motion. Firstly, we learn the human motion model in low-dimensional latent space using nonlinear dimension reduction technique charting based on hierarchical strategy. Increased dependence provision is carried out using hierarchy strategic measures in charting, which improves accuracy in higher flexibility and adaptation. Then we applied QPSO algorithm to estimate the human poses in low-dimensional latent space. Preliminary experimental tracking results show that our approach is able to give good accuracy as compared to conventional state-of-the-arts methods.